COMPMID-718 : Winograd: add validate method and tests

Validate methods added to Winograd kernels and function.
Renamed validation test suit

Change-Id: I0a88df436aff0bbaf4fd82213eeda089b87ac5bf
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/127781
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

6 files changed, 397 insertions, 9 deletions

diff --git a/arm_compute/core/Error.h b/arm_compute/core/Error.h
index 590da9b58e..3635e93244 100644
--- a/arm_compute/core/Error.h
+++ b/arm_compute/core/Error.h
@@ -175,6 +175,16 @@ Status create_error(ErrorCode error_code, const char *function, const char *file
  */
 #define ARM_COMPUTE_CREATE_ERROR_LOC(error_code, func, file, line, ...) ::arm_compute::create_error(error_code, func, file, line, __VA_ARGS__) // NOLINT
 
+/** An error is returned with the given description.
+ *
+ * @param[in] ... Error description message.
+ */
+#define ARM_COMPUTE_RETURN_ERROR_MSG(...)                                                    \
+    do                                                                                       \
+    {                                                                                        \
+        return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, __VA_ARGS__); \
+    } while(false)
+
 /** Checks if a status contains an error and returns it
  *
  * @param[in] status Status value to check
diff --git a/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h b/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h
index 2f44d19b4f..69d8cc6851 100644
--- a/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h
@@ -124,7 +124,7 @@ public:
 
     /** Configure the output transform kernel.
      *
-     * @param[in]  input         Input tensor data
+     * @param[in]  input         Input tensor data. Data types supported: F32.
      * @param[in]  n_batches     Number of batches in input tensor.
      * @param[in]  n_rows        Number of rows in input tensor.
      * @param[in]  n_cols        Number of columns in input tensor.
@@ -152,6 +152,17 @@ public:
     /** Winograd convolution kernel */
     using WinogradConv = typename WinogradBase::template Convolution<T, T>;
 
+    /** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformInputKernel
+     *
+     * @param[in] input       First tensor input info. Data types supported: F32.
+     * @param[in] output      Output tensor info. Data types supported: same as @p input.
+     * @param[in] conv_info   Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported.
+     * @param[in] kernel_dims Kernel dimensions. Currently only 3x3 and 5x5 kernels are supported
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims);
+
 private:
     using InputTransform = typename WinogradBase::template InputTransform<T>;
     std::unique_ptr<InputTransform> _transform;
@@ -301,6 +312,19 @@ public:
     void run(const Window &window, const ThreadInfo &info) override;
     bool is_parallelisable() const override;
 
+    /** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformOutputKernel
+     *
+     * @param[in]  input                 Source tensor with shape [C, N, 16, batches] or [C, N, 36, batches]. Data types supported: F32.
+     * @param[in]  bias                  Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input
+     * @param[out] output                Destination tensor with shape [output_convolved_dims.width, output_convolved_dims.height, C, batches]. Data type supported: same as @p input
+     * @param[in]  kernel_dims           Kernel dimensions (Width and height). Currently only supported 3x3 and 5x5 kernels
+     * @param[in]  output_convolved_dims Output dimensions after the convolution (Width and height)
+     * @param[in]  num_tiles             Number of tiles of size 2x2 or 4x4 in the output tensor along the X and Y direction
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const Size2D &kernel_dims, const Size2D &output_convolved_dims, const Size2D &num_tiles);
+
 private:
     using WinogradBase    = winograd::WinogradGEMM<OutputTileRows, OutputTileCols, KernelRows, KernelCols>;
     using WinogradConv    = typename WinogradBase::template Convolution<T, T>;
@@ -366,6 +390,17 @@ public:
         return "NEWinogradLayerTransformWeightsKernel";
     }
 
+    /** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformWeightsKernel
+     *
+     * @param[in] input       Source tensor info. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout).
+     *                        kernel_x must be 3 and equal to kernel_y. Data types supported: F32.
+     * @param[in] output      Destination tensor info. The output is a 3D tensor with dimensions [OFM, IFM, 16] or [OFM, IFM, 36]. Data type supported: same as @p input
+     * @param[in] output_tile Output tile. Currently only 2x2 and 4x4 tiles are supported.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &output_tile);
+
     // Inherited methods overridden:
     void configure(const ITensor *weights_hwio, T *const output, const int matrix_stride, const int n_output_channels, const int n_input_channels) override;
     unsigned int get_weight_storage_size(int n_output_channels, int n_input_channels) const override;
@@ -496,6 +531,21 @@ public:
 
     void run(const Window &window, const ThreadInfo &info) override;
 
+    /** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerBatchedGEMMKernel.
+     *
+     * @param[in]  a         First input tensor  (Matrix or Vector A). Data types supported: F32
+     * @param[in]  b         Second input tensor (Matrix B). Data type supported: same as @p a.
+     * @param[in]  c         Third input tensor  (Matrix C). It can be a nullptr if just the multiplication between @p a and @p b is needed. Data type supported: same as @p a.
+     * @param[out] output    Output tensor. Data type supported: same as @p a
+     * @param[in]  alpha     Weight of the matrix product
+     * @param[in]  beta      Weight of matrix C
+     * @param[in]  gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and
+     *                       if the reshape of matrix B should happen only for the first run
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensor *c, const ITensorInfo *output, const float alpha, const float beta, const GEMMInfo &gemm_info = GEMMInfo());
+
 private:
     static const int           _output_tile_rows = OutputTileRows;
     static const int           _output_tile_cols = OutputTileCols;
diff --git a/arm_compute/runtime/NEON/functions/NEWinogradLayer.h b/arm_compute/runtime/NEON/functions/NEWinogradLayer.h
index 61a4caae3a..27b1e84201 100644
--- a/arm_compute/runtime/NEON/functions/NEWinogradLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEWinogradLayer.h
@@ -57,7 +57,7 @@ public:
      *                       while every optional dimension from 4 and above represent a batch of inputs.
      *                       Data types supported: F32.
      * @param[in]  weights   Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input.
-     *                       Currently only 3x3 kernels are supported.
+     *                       Currently only 3x3 and 5x5 kernels are supported.
      * @param[in]  biases    Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights.
      * @param[out] output    Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
      *                       Data types supported: Same as @p input.
@@ -75,15 +75,17 @@ public:
      *                      while every optional dimension from 4 and above represent a batch of inputs.
      *                      Data types supported: F32.
      * @param[in] weights   Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
-     *                      Currently only 3x3 kernels are supported.
+     *                      Currently only 3x3 and 5x5 kernels are supported.
      * @param[in] biases    Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights.
      * @param[in] output    Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
      *                      Data types supported: Same as @p input.
      * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported.
+     * @param[in] act_info  (Optional) Activation layer information in case of a fused activation.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                           const ActivationLayerInfo &act_info = ActivationLayerInfo());
 
     /** Prevent instances of this class from being copied (As this class contains pointers) */
     NEWinogradLayer(const NEWinogradLayer &) = delete;
diff --git a/src/core/NEON/kernels/NEWinogradLayerKernel.cpp b/src/core/NEON/kernels/NEWinogradLayerKernel.cpp
index fcd1594601..026a6f1be2 100644
--- a/src/core/NEON/kernels/NEWinogradLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEWinogradLayerKernel.cpp
@@ -23,15 +23,202 @@
  */
 #include "arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h"
 
+#include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
 #include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "support/ToolchainSupport.h"
 
 namespace arm_compute
 {
 //Batched Gemms
+
+namespace
+{
+Status validate_arguments_winograd_gemm(const ITensorInfo *a, const ITensorInfo *b, const ITensor *c, const ITensorInfo *output, const float alpha, const float beta,
+                                        const GEMMInfo &gemm_info = GEMMInfo())
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(a);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(b);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported");
+
+    if(c != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, c->info());
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(1) != c->info()->dimension(1), "The matrix C must have the same number of rows as the matrix A");
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(b->dimension(0) != c->info()->dimension(0), "The matrix C must have the same number of columns as the matrix B");
+    }
+
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(b->dimension(0) != output->dimension(0), "The output matrix must have the same number of columns as the matrix B");
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(1) != output->dimension(1), "The output matrix must have the same number of rows as the matrix A");
+        ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() != a->num_dimensions());
+    }
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(0) != b->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
+    ARM_COMPUTE_UNUSED(alpha, beta);
+    return Status{};
+}
+
+Status validate_arguments_winograd_weight_trans(const ITensorInfo *input, const ITensorInfo *output, const Size2D &output_tile)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+
+    const size_t idx_width  = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
+    const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_width) != 3 && input->dimension(idx_width) != 5);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_width) != input->dimension(idx_height));
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
+    ARM_COMPUTE_RETURN_ERROR_ON(output_tile != Size2D(2U, 2U) && output_tile != Size2D(4U, 4U));
+
+    // Checks performed when output is configured
+    if(output->total_size() != 0)
+    {
+        const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_winograd_filter_transform_shape(*input, output_tile));
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window_winograd_weight_trans(ITensorInfo *input, ITensorInfo *output, const Size2D &output_tile, const Size2D &kernel_dims)
+{
+    ARM_COMPUTE_UNUSED(output_tile);
+
+    // Output tensor auto inizialitation if not yet initialized
+    auto_init_if_empty(*output, input->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_winograd_filter_transform_shape(*input, output_tile)));
+
+    unsigned int num_elems_processed_per_iteration_x = kernel_dims.width;
+    unsigned int num_elems_processed_per_iteration_y = kernel_dims.height;
+
+    Window win            = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+    bool   window_changed = false;
+
+    AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
+    AccessWindowStatic    output_access(output, 0, 0, output->dimension(0), output->dimension(1));
+    window_changed = update_window_and_padding(win, input_access, output_access);
+    output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
+
+    Window win_collapsed = win.collapse(win, Window::DimZ);
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+
+    return std::make_pair(err, win_collapsed);
+}
+
+Status validate_arguments_winograd_input_trans(const ITensorInfo *input, const ITensorInfo *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd input transform only supports unit strides");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != 3U && kernel_dims.width != 5U), "Winograd input transform only supports 3x3 and 5x5 kernels");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != kernel_dims.height), "Winograd input transform only supports 3x3 and 5x5 kernels");
+
+    // Validate configured output
+    if(output->total_size() != 0)
+    {
+        const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, conv_info, kernel_dims);
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window_winograd_input_trans(ITensorInfo *input, ITensorInfo *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims,
+                                                                             const Size2D &tile_dims)
+{
+    const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, conv_info, kernel_dims);
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape));
+
+    unsigned int num_elems_read_per_iteration_x = (tile_dims.width + kernel_dims.width - 1);
+    unsigned int num_elems_read_per_iteration_y = (tile_dims.height + kernel_dims.height - 1);
+
+    Window win = calculate_max_window(*input, Steps(1, 1));
+
+    AccessWindowRectangle input_access(input, -conv_info.pad_left(), -conv_info.pad_top(), num_elems_read_per_iteration_x, num_elems_read_per_iteration_y);
+
+    bool window_changed = update_window_and_padding(win, input_access);
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+
+Status validate_arguments_winograd_output_trans(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const Size2D &kernel_dims, const Size2D &output_convolved_dims,
+                                                const Size2D &num_tiles)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != num_tiles.area());
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != 3U && kernel_dims.width != 5U), "Winograd output transform only supports 3x3 and 5x5 kernels");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != kernel_dims.height), "Winograd output transform only supports 3x3 and 5x5 kernels");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(((input->dimension(2) != size_t(16U)) && (input->dimension(2) != size_t(36U))), "Only 2x2 and 4x4 output tile is supported");
+    ARM_COMPUTE_UNUSED(kernel_dims);
+    if(bias != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
+        ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
+        ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() != size_t(1));
+    }
+
+    // Checks performed when output is configured
+    if(output->total_size() != 0)
+    {
+        const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_winograd_output_transform_shape(*input, output_convolved_dims, DataLayout::NCHW));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window_winograd_output_trans(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output, const Size2D &output_convolved_dims)
+{
+    // Output tensor auto initialization if not yet initialized
+    auto_init_if_empty(*output, input->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_winograd_output_transform_shape(*input, output_convolved_dims, DataLayout::NCHW)));
+
+    constexpr unsigned int num_elems_processed_per_iteration = 1;
+
+    Window win            = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+    bool   window_changed = false;
+
+    AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration, num_elems_processed_per_iteration);
+    AccessWindowStatic    output_access(output, 0, 0, ceil_to_multiple(output->dimension(0), 2), ceil_to_multiple(output->dimension(1), 2));
+
+    if(bias != nullptr)
+    {
+        AccessWindowStatic bias_access(bias, 0, 0, bias->dimension(0), bias->dimension(1));
+        window_changed = update_window_and_padding(win, input_access, bias_access, output_access);
+    }
+    else
+    {
+        window_changed = update_window_and_padding(win, input_access, output_access);
+    }
+    output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
 template <typename TIn, typename TOut, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
 NEWinogradLayerBatchedGEMMKernel<TIn, TOut, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::NEWinogradLayerBatchedGEMMKernel()
     : _gemms()
@@ -93,6 +280,14 @@ int NEWinogradLayerBatchedGEMMKernel<TIn, TOut, OutputTileRows, OutputTileCols,
     return WinogradConv::N_BLOCK;
 }
 
+template <typename TIn, typename TOut, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+Status NEWinogradLayerBatchedGEMMKernel<TIn, TOut, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensor *c,
+                                                                                                                     const ITensorInfo *output, const float alpha, const float beta, const GEMMInfo &gemm_info)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_gemm(a, b, c, output, alpha, beta, gemm_info));
+    return Status{};
+}
+
 template class NEWinogradLayerBatchedGEMMKernel<float, float, 2, 2, 3, 3>;
 template class NEWinogradLayerBatchedGEMMKernel<float, float, 2, 2, 5, 5>;
 
@@ -152,6 +347,14 @@ bool NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTileCols, Ke
     return false;
 }
 
+template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+Status NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &output_tile)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_weight_trans(input, output, output_tile));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_winograd_weight_trans(input->clone().get(), output->clone().get(), output_tile, Size2D(KernelRows, KernelCols)).first);
+    return Status{};
+}
+
 template class NEWinogradLayerTransformWeightsKernel<float, 2, 2, 3, 3>;
 template class NEWinogradLayerTransformWeightsKernel<float, 2, 2, 5, 5>;
 
@@ -222,6 +425,16 @@ bool NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, Kern
     return false;
 }
 
+template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+Status NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::validate(const ITensorInfo *input, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                                                                                                                const Size2D &kernel_dims)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_input_trans(input, output, conv_info, kernel_dims));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_winograd_input_trans(input->clone().get(), output->clone().get(), conv_info, kernel_dims, Size2D(OutputTileRows, OutputTileCols)).first);
+
+    return Status{};
+}
+
 template class NEWinogradLayerTransformInputKernel<float, 2, 2, 3, 3>;
 template class NEWinogradLayerTransformInputKernel<float, 2, 2, 5, 5>;
 
@@ -318,6 +531,19 @@ bool NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, Ker
     return false;
 }
 
+template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+Status NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
+                                                                                                                 const Size2D &kernel_dims, const Size2D &output_convolved_dims,
+                                                                                                                 const Size2D &num_tiles)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_output_trans(input, (bias != nullptr ? bias->clone().get() : nullptr), output, kernel_dims, output_convolved_dims, num_tiles));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_winograd_output_trans(input->clone().get(), (bias != nullptr ? bias->clone().get() : nullptr), output->clone().get(),
+                                                                                    output_convolved_dims)
+                                .first);
+
+    return Status{};
+}
+
 template class NEWinogradLayerTransformOutputKernel<float, 2, 2, 3, 3>;
 template class NEWinogradLayerTransformOutputKernel<float, 2, 2, 5, 5>;
 
diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
index b0603e92d2..61ea2db15b 100644
--- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
@@ -83,7 +83,7 @@ Status NEConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo
     {
         case ConvolutionMethod::WINOGRAD:
             //Validate Winograd
-            NEWinogradLayer::validate(input, weights, biases, output, conv_info);
+            NEWinogradLayer::validate(input, weights, biases, output, conv_info, act_info);
             break;
         case ConvolutionMethod::GEMM:
             //Validate Gemm-based Convolution
diff --git a/src/runtime/NEON/functions/NEWinogradLayer.cpp b/src/runtime/NEON/functions/NEWinogradLayer.cpp
index 126be46b2e..7f4761020c 100644
--- a/src/runtime/NEON/functions/NEWinogradLayer.cpp
+++ b/src/runtime/NEON/functions/NEWinogradLayer.cpp
@@ -26,6 +26,8 @@
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 #include "support/ToolchainSupport.h"
 
@@ -51,6 +53,9 @@ namespace
 {
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info)
 {
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(weights);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != 3 && weights->dimension(0) != 5, "Only 3 and 5 kernels are supported");
@@ -69,7 +74,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(stride_y != 1 || stride_x != 1, "Winograd layer only supports unit strides.");
 
     ARM_COMPUTE_UNUSED(output);
-
     return Status{};
 }
 } //namespace
@@ -258,11 +262,107 @@ void NEWinogradLayer::run()
     _memory_group.release();
 }
 
-Status NEWinogradLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info)
+Status NEWinogradLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                                 const ActivationLayerInfo &act_info)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_RETURN_ERROR_ON(validate_arguments(input, weights, biases, output, conv_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info));
+
+    // Get indices for the width and height
+    const size_t idx_width  = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
+    const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
+
+    // Kernel size
+    const unsigned int kernel_w = weights->tensor_shape()[idx_width];
+    const unsigned int kernel_h = weights->tensor_shape()[idx_height];
+
+    // Number of tiles along the X and Y direction
+    const unsigned int num_tiles_x = std::ceil((input->tensor_shape().x() - (kernel_w - 1) + conv_info.pad_left() + conv_info.pad_right()) / 2.f);
+    const unsigned int num_tiles_y = std::ceil((input->tensor_shape().y() - (kernel_h - 1) + conv_info.pad_top() + conv_info.pad_bottom()) / 2.f);
+
+    // Compute output shape
+    const TensorShape output_convolved_shape = misc::shape_calculator::compute_deep_convolution_shape(*input, *weights, conv_info);
 
+    // Validate input transform
+    const TensorShape input0_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, conv_info, Size2D(kernel_w, kernel_h));
+    const TensorInfo  input0       = input->clone()->set_tensor_shape(input0_shape);
+    switch(weights->dimension(0))
+    {
+        case 3:
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel<float, 2, 2, 3, 3>::validate(input, &input0, conv_info, Size2D(kernel_w, kernel_h))));
+            break;
+        }
+        case 5:
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel<float, 2, 2, 5, 5>::validate(input, &input0, conv_info, Size2D(kernel_w, kernel_h))));
+            break;
+        }
+        default:
+        {
+            ARM_COMPUTE_RETURN_ERROR_MSG("Only 3x3 and 5x5 kernels supported.");
+            break;
+        }
+    }
+    // Validate filter transform
+    const TensorShape input1_shape = misc::shape_calculator::compute_winograd_filter_transform_shape(*weights, Size2D(2U, 2U));
+    const TensorInfo  input1       = weights->clone()->set_tensor_shape(input1_shape);
+
+    switch(weights->dimension(0))
+    {
+        case 3:
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel<float, 2, 2, 3, 3>::validate(weights, &input1, Size2D(2U, 2U))));
+            break;
+        }
+        case 5:
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel<float, 2, 2, 5, 5>::validate(weights, &input1, Size2D(2U, 2U))));
+            break;
+        }
+        default:
+        {
+            ARM_COMPUTE_RETURN_ERROR_MSG("Only 3x3 and 5x5 kernels supported.");
+            break;
+        }
+    }
+    // Validate batched matrix multiply
+    TensorShape batched_mm_output_shape = input0.tensor_shape();
+    batched_mm_output_shape[0]          = input1.tensor_shape()[0];
+    const TensorInfo batched_mm_output  = input0.clone()->set_tensor_shape(batched_mm_output_shape);
+    switch(weights->dimension(0))
+    {
+        case 3:
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerBatchedGEMMKernel<float, float, 2, 2, 3, 3>::validate(&input0, &input1, nullptr, &batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false,
+                                                                                                              true /* Reshape weights only for the first run*/))));
+            // Validate output transform
+            ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel<float, 2, 2, 3, 3>::validate(&batched_mm_output, biases, output, Size2D(kernel_w, kernel_h), Size2D(output_convolved_shape[idx_width],
+                                                                                                           output_convolved_shape[idx_height]),
+                                                                                                           Size2D(num_tiles_x, num_tiles_y))));
+            break;
+        }
+        case 5:
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerBatchedGEMMKernel<float, float, 2, 2, 5, 5>::validate(&input0, &input1, nullptr, &batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false,
+                                                                                                              true /* Reshape weights only for the first run*/))));
+            // Validate output transform
+            ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel<float, 2, 2, 5, 5>::validate(&batched_mm_output, biases, output, Size2D(kernel_w, kernel_h), Size2D(output_convolved_shape[idx_width],
+                                                                                                           output_convolved_shape[idx_height]),
+                                                                                                           Size2D(num_tiles_x, num_tiles_y))));
+            break;
+        }
+        default:
+        {
+            ARM_COMPUTE_RETURN_ERROR_MSG("Only 3x3 and 5x5 kernels supported.");
+            break;
+        }
+    }
+
+    // Validate Activation Layer
+    if(act_info.enabled())
+    {
+        NEActivationLayer::validate(output, nullptr, act_info);
+    }
     return Status{};
 }